MedLSAM: Localize and segment anything model for 3D CT images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Foundation models show promise in medical image analysis but lack specific 3D localization capabilities.
• Accurate localization of anatomical structures is crucial for various medical imaging applications.

Purpose of the Study:

• To introduce MedLAM, a 3D medical foundation localization model for precise anatomical part identification.
• To develop MedLSAM by integrating MedLAM with the Segment Anything Model (SAM) to reduce annotation efforts.
• To evaluate the performance of MedLAM and MedLSAM in localizing and segmenting anatomical structures.

Main Methods:

• MedLAM utilizes two self-supervision tasks: unified anatomical mapping (UAM) and multi-scale similarity (MSS).
• A dataset of 14,012 CT scans was used for training and validation.
• MedLSAM integrates MedLAM with SAM, requiring minimal point annotations for localization and segmentation.

Main Results:

• MedLAM achieved performance comparable to fully supervised models in localizing anatomical structures using few template scans.
• MedLSAM demonstrated performance on par with SAM using manual prompts, significantly reducing annotation requirements.
• Experiments were conducted on two 3D datasets covering 38 distinct organs.

Conclusions:

• MedLAM effectively localizes anatomical structures with minimal template data.
• MedLSAM significantly lowers the annotation burden for 3D medical image segmentation tasks.
• MedLAM offers potential for integration with future 3D SAM models to improve segmentation accuracy.